Autofluorescence Imaging of 3D Tumor–Macrophage Microscale Cultures Resolves Spatial and Temporal Dynamics of Macrophage Metabolism

Immune cells are a critical component of the tumor microenvironment (TME) and regulate tumor initiation, progression, and therapeutic response (1). The phenotype and function of immune cells surrounding or infiltrating the tumor are crucial to tumor fate (1). Macrophages are a subset of immune cells with diverse function that are abundant within numerous tumor types, especially breast and pancreatic tumors (1, 2). This prevalence suggests that macrophages may serve as important immune mediators of tumor behavior, yet macrophage activity is complex and not well understood. For example, multiple macrophage phenotypes have been identified, each with unique functional and stimulatory activity that influence the surrounding cellular environment (3, 4). Two well-characterized classes of macrophages, M1-like and M2-like macrophages, exhibit distinct responses to tumors (4). M1-like macrophages promote antitumor activity through apoptotic or phagocytic signaling, secondary immune stimulation via cytokine secretion and antigen presentation, and nutrient deprivation of tumors (4, 5). M2-like macrophages support tumor growth by suppressing immune recognition and promoting cancer cell proliferation, motility, and vessel growth (4, 5). However, intermediate macrophage states have also been observed within tumors, suggesting a more complex model of macrophage heterogeneity than this M1–M2 dichotomy (6). This heterogeneity within macrophage populations presents a substantial challenge for effective treatments, emphasizing the need to characterize cell-level macrophage function within the 3D TME.

Cell function has been correlated with metabolism across numerous cell types, including macrophages (4, 6, 7). Therefore, monitoring metabolic preferences within macrophages could reveal heterogeneity in macrophage phenotype and function. Metabolic demands greatly differ between M1-like and M2-like macrophage classes (4). M1-like macrophages upregulate glycolysis to sustain viability, induce tumor cell death, and recruit additional monocytes to the tumor. This increases the production of the metabolic coenzyme, NADH, via glycolysis, and sustains the viability of M1-like macrophages within tumors (6, 7). Conversely, elevated fatty acid oxidation and oxidative phosphorylation in M2-like macrophages support angiogenesis and tumor growth (6–8). Microenvironmental changes can regulate these distinct metabolic demands and produce diverse macrophage phenotypes (6, 9). This highlights a continuum of metabolic activity related to macrophage function that is provoked by complex stimuli from the TME (5, 6). Therefore, monitoring macrophage metabolic heterogeneity may improve understanding of tumor–macrophage cross-talk.

3D tumor models provide reliable portrayal of tumor structure and TME conditions that influence immune cell function. The key advantages of microscale models include controllable features that can be easily tuned to simulate the TME and designs enabling various combinations of culture conditions that support high-throughput assessment (10). These microscale platforms also provide a simple approach to recapitulate paracrine signaling, protein–protein interactions, and in vivo tumor-associated environmental conditions (e.g., hypoxia and nutrient availability) that regulate both macrophage phenotype and function (10–12). These representative 3D tumor models provide attractive systems to evaluate macrophage heterogeneity in response to spatiotemporal changes in the microenvironment.

Limitations of destructive, population-level functional assays (e.g., flow cytometry and ELISA) and low-resolution whole-body imaging (e.g., PET/CT and MRI) prohibit monitoring spatiotemporal macrophage heterogeneity within the complex 3D TME (13, 14). Noninvasive autofluorescence imaging has previously quantified cellular heterogeneity in vivo and in 3D cultures using label-free two-photon microscopy (15–19). Specifically, autofluorescence imaging can quantify the endogenous fluorescence of metabolic coenzymes, NADH and FAD, which are both used across several cellular metabolic processes (20). NADH and NADPH have overlapping fluorescence properties, and are collectively referred to as NAD(P)H (21). Fluorescence intensity measurements can inform on the intracellular concentrations of NAD(P)H and FAD. The optical redox ratio, defined as the ratio of NAD(P)H intensity to FAD intensity, provides a measure of the relative oxidation–reduction state of the cell (15, 20). Fluorescence lifetime imaging microscopy (FLIM) of NAD(P)H and FAD provides additional information specific to protein binding activity (19). Lifetime measurements can distinguish between free and protein-bound forms of NAD(P)H and FAD, characterized by distinct molecular conformations that affect fluorescence quenching (19). Previous studies have shown that metabolic autofluorescence imaging detects spatial and temporal changes in stromal cells across in vivo and 3D in vitro models (17, 22–24). This study aimed to validate whether autofluorescence imaging of NAD(P)H and FAD can monitor changes in macrophage metabolism with polarization and migration. A novel technology combining metabolic autofluorescence imaging and microscale 3D models was established to quantify metabolic activity and visualize macrophage heterogeneity within the 3D TME using primary human cancer, human cell lines, and mouse cell lines. These technologies could provide valuable insights for future studies of tumor-associated macrophage function.

RAW264.7 murine macrophages (ATCC, TIB-71) were maintained in culture medium composed from standard RPMI1640 (Gibco), 10% FBS, and 1% penicillin–streptomycin. RAW macrophages were seeded in 35-mm glass bottom dishes (MatTek) for imaging experiments. All imaging samples were plated at a density of 1 × 10⁵ cells per dish and incubated at 37°C and 5% CO₂, for 24 hours to allow cell adhesion. Separate cultures of macrophages polarized with M1-like (IFNγ) or M2-like (IL4/IL13) cytokines were generated by standard media substitution with 2 mL cytokine-supplemented media and incubated between 24 and 72 hours. Media for M(IFNγ) stimulation consisted of RPMI1640 supplemented with 10 ng/mL IFNγ (R&D Systems), and media for M(IL4/IL13) stimulation consisted of RPMI1640 with 20 ng/mL IL4 (Invitrogen) and 20 ng/mL IL13 (Gibco).

Description of the Stacks microfluidic platform is described in detail previously (11). Twenty-four-well polystyrene microdevices (dimensions: 75 mm × 50 mm) were fabricated by injection molding and sterilized via sonication in isopropanol. Collagen hydrogels (2 mg/mL) were prepared and a 4.5 μL volume was suspended within each microwell of the device (thickness ∼1.2 mm), creating an open-air culture system compatible for two-photon imaging. The hydrogel was generated from a mixture of 6 μL 10X PBS, 14 μL sterile water, 6 μL NaOH (Sigma), 160 μL Bovine Collagen Type I (PureCol, Advanced BioMatrix), 3 μL fibronectin (Sigma), and 50 μL cell type–specific medium [RPMI1640 for mouse and DMEM/complete DMEM/F12 (Gibco) for human]. Collagen hydrogels were polymerized by incubation at 37°C for at least 6 hours, prior to sequential seeding of breast carcinoma cells and macrophages (∼1,000 cells/μL each) on the opposing ends of the collagen layer.

The 3D microfluidic platform was used for cocultures of primary and immortalized macrophages and breast carcinoma of both mouse and human origin. Polyoma-Middle T Virus (PyVMT; established from FVB MMTV-PyVmT mouse tumors) and MDA-MB-231 (ATCC, HTB-26) human breast cancer cells were cultured in medium composed from standard DMEM (Gibco), 10% FBS, and 1% penicillin–streptomycin. RAW264.7 macrophages (ATCC, TIB-71) and THP-1 human monocytes (ATCC, TIB-202) were cultured in RPMI1640 medium supplemented with 10% FBS and 1% penicillin–streptomycin. All immortalized lines were used for less than 15 passages from thawing. Patient-derived invasive breast carcinoma cells (IDC; 171881-019-R-J1-PDC, passages 7–12) were obtained upon request from the NCI Patient-Derived Models Repository (NCI PDMR; ref. 25). Culture of patient-derived cells required specialized media consisting of 1X advanced DMEM/F12 (Gibco) supplemented with 5% FBS, 1.1 μmol/L Hydrocortisone (Sigma), 1.61 nmol/L EGF Recombinant Human Protein (Invitrogen), 0.178 mmol/L Adenine (Sigma), 1% penicillin–streptomycin, 2 mmol/L l-Glutamine (Invitrogen), and 0.01 mmol/L Y-27632 dihydrochloride (Tocris; ref. 25). All cell lines were tested negative for Mycoplasma by the University of Wisconsin-Madison (Madison, WI) or NCI PDMR and verified with short-tandem repeat profiling by the ATCC and NCI PDMR. 3D microwell conditions included macrophages incubated with and without tumor cells over 72 hours. For gene expression analysis, tumor cells and macrophages were cultured on separate Stacks layers to prevent cell type cross-contamination. An additional Stacks layer with wells only containing 10 μL culture medium was placed on top of each Stacks layer with cells. These layers were then overlaid forming a four-layer Stacks setup (media layer–macrophage layer–media layer–tumor layer). Media were changed daily by aspirating old media from each microwell and replacing with 10 μL of fresh culture media during incubation between imaging experiments.

FLIM images were acquired with an Ultima two-photon imaging system (Bruker) composed of an ultrafast tunable laser source (Insight DS+, Spectra Physics) coupled to a Nikon Ti-E inverted microscope equipped with time-correlated single-photon counting electronics (SPC 150, Becker & Hickl GmbH). The ultrafast tunable laser source enabled excitation of NAD(P)H (750 nm) and FAD (890 nm) fluorescence. For 2D macrophage imaging, NAD(P)H and FAD fluorescence were excited sequentially, and their respective emission was isolated using 440/80 and 550/100 bandpass filters (Chroma). Laser power at the sample for NAD(P)H and FAD excitation was approximately 11.3 and 11.5 mW, respectively. Simultaneous excitation of NAD(P)H (750 nm) and FAD (895 nm) was performed for all 3D samples by wavelength mixing of 750 and 1,041 nm laser lines to virtually generate a 895 nm excitation effect as described previously (26, 27). Wavelength mixing was implemented by delaying the tunable laser line and matching the collimation of both the fixed 1,041 nm and tunable lasers with a single-beam telescope to align both laser lines spatially and temporally within the focal volume (Supplementary Fig. S2A). Efficient 895 nm excitation was confirmed with excitation of fixed GFP-labeled cells across a range of 750 and 1,041 nm laser powers (0–1.5 mW) relative to single-color 895 nm excitation, supporting minimal excitation at the single 750 or 1,041 nm wavelengths compared with dual excitation. Colocalization of GFP signal with one- (1,041 nm) and two-laser (750 and 1,041 nm) excitation was also used to regularly correct alignment of the laser lines. For wavelength mixing experiments, all samples were illuminated through a 40×/1.15 NA objective (Nikon) with an optical zoom parameter of 2, decreasing the scan angle and resulting field of view to capture fluorescence emission only from the lateral area of overlap between the two laser lines. Approximate laser power at the sample during wavelength mixing was 10.7 mW at 750 nm and 16.3 mW at 1,041 nm. NAD(P)H and FAD emissions during wavelength mixing were isolated using a 466/40 and a 540/24 nm bandpass filter (Chroma), respectively. Wavelength mixing results were further validated by comparing two-color sequential imaging and wavelength mixing autofluorescence trends in standard culture of a breast carcinoma cell line (BT474, ATCC), as well as following perturbation with 4 mmol/L sodium cyanide (NaCN) for 5 minutes in BT474 monolayers (27). Emission was detected with GaAsP photomultiplier tubes (Hamamatsu, 7422PA-40). Per-pixel lifetime decay curves were collected by scanning the field of view (256 × 256 pixels; 300 μm × 300 μm) over 60 seconds with a 4.8 microsecond pixel dwell time. To prevent photobleaching within the sample, photon count rates were maintained at approximately 1–3 × 10⁵ photons/second. The instrument response function was generated from second harmonic signal of urea crystals excited at 900 nm, with a full width at half maximum of 240 picoseconds. A fluorescence lifetime standard measurement was collected daily by imaging a YG fluorescence bead (Polysciences, Inc.), measured as 2.1 ± 0.09 nanoseconds consistent with reported lifetime values (28). Autofluorescence images were captured for each 2D polarization condition at 24, 48, and 72 hours poststimulation, with a minimum of four representative fields of view per sample (∼1,000–2,000 cells). For 3D cultures, autofluorescence lifetime volumes were acquired across 2–3 microwells for each macrophage monoculture and coculture condition for at least three separate devices (i.e., 6–9 microwells/condition). Image volumes captured images at 3-μm z-steps starting at the macrophage seeding plane and ending at the leading edge of the macrophage layer (∼1,000–3,000 cells). Z-stack image depths ranged from 24 to 170 μm.

Per-pixel fluorescence lifetimes of free and protein-bound NAD(P)H and FAD were calculated from fitting fluorescence decays to the following biexponential model: |${\rm{I}}({\rm{t}})\ = \ {\alpha _1}{e}^{{ - t}/{\tau _1}}$| + |{\alpha _2}{e}^{{ - t}/{\tau _2}}$| + C (19). Fluorescence intensity images were generated by integrating photon counts over the per-pixel fluorescence decays. The per-pixel ratio of NAD(P)H fluorescence intensity to FAD intensity was calculated to determine optical redox ratio. A customized CellProfiler pipeline was used to segment individual cell cytoplasms (19). Cytoplasm masks were applied to all images to determine single-cell redox ratio and NAD(P)H and FAD fluorescence lifetime variables. Fluorescence lifetime variables consist of the mean lifetime (τ_m = τ₁α₁ + τ₂α₂), free- and protein-bound lifetime components (τ₁ and τ₂), and their fractional contributions (α₁ and α₂; where α₁ + α₂ = 1) for each cell cytoplasm. A total of 11 variables were analyzed for each cell cytoplasm: NAD(P)H τ_m, FAD τ_m, NAD(P)H τ₁, FAD τ₁, NAD(P)H τ₂, FAD τ₂, NAD(P)H α₁, FAD α₁, NAD(P)H intensity, FAD intensity, and optical redox ratio.

To validate that metabolic changes were specific to macrophage polarization, 2D-polarized macrophages were treated with a panel of metabolic inhibitors with known activity and imaged posttreatment. RAW264.7 macrophages were stimulated to glycolytic M(IFNγ) or oxidative M(IL4/IL13) phenotype for 72 hours (29, 30). Next, glycolysis was inhibited with media + 10 mmol/L 2-deoxyglucose (2DG; Sigma), fatty acid oxidation was inhibited with media + 100 nmol/L etomoxir (Sigma), or oxidative phosphorylation was inhibited with media + 4 mmol/L NaCN (Sigma; ref. 31). Autofluorescence imaging was performed immediately before adding inhibitor and after treatment for 5 minutes (NaCN), 1 hour (2DG), or 24 hours (etomoxir).

Macrophage migration was plotted as the percentage of macrophages within each z-plane versus z-plane distance from the macrophage seeding plane. Specifically, the percent of macrophages at each z-plane was calculated as the number of macrophages at each z-plane divided by the total number of macrophages in each microwell. This percent of macrophages for each z-plane was plotted versus z-plane distance from the macrophage seeding plane for histograms of macrophage migration distance, to further define migratory behavior (i.e., “passive” and “active” migration). The maximum z-plane distance for the monoculture condition at each timepoint was used as the threshold to discriminate passively (closest to macrophage seeding plane) versus actively migrating (farthest from macrophage seeding plane) macrophages in cocultures. Passive migration was defined as random, undirected migratory activity, while active migration represented a directional chemotactic response (32).

Following autofluorescence imaging, unstimulated and cytokine-stimulated 2D cultures of RAW264.7 macrophages were stained according to the manufacturer's instructions with both PE-conjugated CD86 antibody (Tonbo Biosciences) and FITC-conjugated CD206 antibody (Bio-Rad) to confirm macrophage polarization. Conjugated antibodies were diluted to 1:30 in PBS + goat serum. 2D immunofluorescence images were acquired using the two-photon microscope setup described above. CD206-FITC fluorescence intensity was excited at 890 nm and collected with a 550/100 nm bandpass filter, while CD86-PE fluorescence was excited at 1,050 nm and collected with a 650/47 nm filter.

To assess genetic heterogeneity in macrophage polarization, expression levels of genes specific to broad macrophage classes (i.e., M1-like, M2-like, and mixed phenotypes) were analyzed by qRT-PCR. For 2D cultures, separate cultures of M0, M(IFNγ), and M(IL4/IL13) RAW 264.7 macrophages were maintained in 6-well plates prior to mRNA extraction. For 3D cultures, RAW264.7 macrophages or THP-1 monocytes were cultured alone or with breast cancer cells in the 3D Stacks microfluidic platform. For all experiments, mRNA was extracted at 24, 48, and 72 hours using a Dynabeads mRNA DIRECT Purification Kit (Thermo Fisher Scientific, 61012). Purified mRNA was quantified using a Qubit Fluorometer (Thermo Fisher Scientific) and a Qubit RNA BR Assay Kit (Thermo Fisher Scientific, Q10210). mRNA was reverse transcribed using RT² PreAMP cDNA Synthesis Kit (Qiagen, 330451). The prepared cDNA was preamplified using the RT² PreAMP Primer Mix for Human and Mouse PCR Array (Qiagen, PBH-181Z). cDNA was analyzed by qRT-PCR using a Qiagen RT2 Profiler Custom Panel (Qiagen, PAHS-181Z, CLAM3313, CLAH36077) following the manufacturer's instructions. Gene lists for mouse and human experiments are reported in Supplementary Table S1.

Population density modeling of single-cell metabolism was used to observe differences in macrophage heterogeneity across 2D-polarized cultures and 3D tumor–macrophage cultures over time. Frequency histograms were generated on the basis of single-cell metabolic autofluorescence measurements across samples. Multiple Gaussian probability density functions were fit to each distribution histogram to identify the presence of subpopulations with distinct metabolic activity (33). Accurate subpopulation identification was evaluated by iteratively increasing the number of fitted Gaussian curves and calculating the Akaike information criterion (AIC) for each iteration (33). The number of fitted Gaussians yielding the lowest AIC value indicated the optimal fit conditions and number of subpopulations per sample (33).

Multi-class random forest classifiers were generated and trained using Python (Python Software Foundation) to distinguish (i) macrophage polarization conditions [M0, M(IFNγ), and M(IL4/IL13)] pooled across the stimulation time course (24, 48, and 72 hours) and passively and actively migrating cells in (ii) mouse PyVMT + RAW264 cocultures and (iii) human IDC + THP cocultures across timepoints. Conditions were pooled across timepoints to allow robust classification of each group without overfitting. Autofluorescence measurements for all 2D macrophage conditions (cell number = 14,504), mouse 3D macrophage conditions (cell number = 360), and human 3D macrophage conditions (cell number = 4,079), respectively, were randomly sampled from each condition and timepoint to generate balanced classification datasets before assignment to training and test sets. Classification accuracy was assessed as a function of varying proportions of training/test data: 12.5%/87.5%, 25%/75%, 37.5%/62.5%, 50%/50%, 62.5%/37.5%, 75%/25%, and 87.5%/12.5%.

Tukey tests for nonparametric, unpaired comparisons were performed to assess differences in autofluorescence variables between 2D cytokine-stimulated cultures and 3D monoculture and coculture conditions over 72 hours. Multiple Student t tests were performed to assess significance of fold change differences for redox ratio changes in 2D macrophage subpopulation analysis and in response to metabolic inhibitors and gene expression changes with respect to M0 conditions for 2D mouse macrophages or monoculture conditions for 3D mouse and human macrophages. Coefficient of variation was calculated to evaluate variability across 2D-polarized mouse cultures and 3D mouse monocultures and cocultures over 72 hours. The nonparametric, squared-ranks test was performed to assess the equality of coefficients of variation between macrophage groups (MATLAB, RRID:SCR_001622, SquaredRanksTest; ref. 34). Metabolic autofluorescence results are represented as line plots, dot plots, or bar graphs plotted in GraphPad (GraphPad Prism, RRID:SCR_002798). Heatmaps of autofluorescence variables with respect to depth were generated in MATLAB.

Autofluorescence imaging can nondestructively monitor live, single-cell metabolism and 3D migration, while 3D microscale cocultures reliably model tumor complexity with selective environmental control and high-throughput capabilities. Specifically, the Stacks 3D microfluidic system enables multi-cellular cultures with minimal biological material, flexible configuration, and well-characterized environmental gradients (11). The Stacks system was combined with metabolic autofluorescence imaging to monitor tumor-mediated changes in macrophage metabolism and migration (Fig. 1A; ref. 11). Microscale 3D cocultures of tumor cells and macrophages were established in the Stacks 3D system using primary human cancer cells, and cell lines from mouse and of human origin (Fig. 1A). A collagen-based extracellular matrix (ECM) was polymerized within each microwell, then monocytes/macrophages were seeded with (coculture) or without (monoculture) breast cancer cells at the opposing end of the ECM layer (Fig. 1A). Both the metabolic autofluorescence and microfluidic culture technologies were validated by imaging macrophages in 2D culture with standard stimulation techniques and in 3D cocultures of macrophages alone in the Stacks 3D system (Fig. 1B). Finally, the metabolic dynamics of tumor-stimulated macrophages were characterized by collecting NAD(P)H and FAD autofluorescence image volumes of monocytes/macrophages migrating across the ECM at 24, 48, and 72 hours (Fig. 1C). Overall, this approach provides unique, quantitative imaging of macrophage spatiotemporal dynamics in the TME that cannot be achieved with conventional assessments that lack single-cell resolution, 3D complexity, and/or nondestructive monitoring capability.

To establish the sensitivity of metabolic autofluorescence imaging to distinct macrophage phenotypes, metabolic autofluorescence measurements were first validated in 2D cultures of RAW264.7 mouse macrophages stimulated with IFNγ [M(IFNγ), antitumor M1-like phenotype], IL4/IL13 [M(IL4/IL13), protumor M2-like phenotype], or without cytokine stimulation (M0, naïve macrophages). Macrophages were maintained in cytokine-supplemented media and imaged over 24–72 hours. Autofluorescence imaging of NAD(P)H and FAD demonstrated time-dependent differences in M(IFNγ) and M(IL4/IL13) macrophages (Fig. 2A). M(IFNγ) macrophages had increased redox ratio compared with M(IL4/IL13) macrophages, with the greatest redox ratio differences at 72 hours postpolarization (Fig. 2A). Individual M(IFNγ) and M(IL4/IL13) macrophages had heterogeneous NAD(P)H and FAD lifetimes across fields of view at all timepoints. A flattened, cuboid structure was observed for M(IFNγ) macrophages, whereas M(IL4/IL13) macrophages extended outward with spindle projections, consistent with reported morphology of polarized macrophages (Fig. 2A; ref. 35). These known subset-specific morphologic changes became more pronounced over time (Fig. 2A). Staining for common M1-like and M2-like surface proteins (CD86 and CD206, respectively) showed increasing CD86 expression in M(IFNγ) macrophages across the 72-hour time course, while M(IL4/IL13) macrophages had stronger CD206 expression (Fig. 2A).

Quantitative autofluorescence measurements were then compared between M0, M(IFNγ), and M(IL4/IL13) macrophages over the 72-hour time course to determine whether metabolic autofluorescence can distinguish macrophage polarization states. Z-score heatmaps of all 11 metabolic variables show the average value of each polarization state with respect to unpolarized, M0 macrophages (Fig. 2B). The full distributions of each metabolic variable are observed in Supplementary Fig. S1. The redox ratio was significantly different (P < 0.0001) between M0, M(IFNγ), and M(IL4/IL13) macrophages at 48 and 72 hours postpolarization (Fig. 2B; Supplementary Fig. S1A). NAD(P)H and FAD lifetimes further demonstrated dynamic metabolic behavior across all three macrophage conditions, consistently exhibiting significant differences 24 hours poststimulation (P < 0.05 and 0.001; Fig. 2B; Supplementary Fig. S1B and S1C). The short (τ₁) and long lifetimes (τ₂) and their relative contribution (α₁) for NAD(P)H and FAD influenced changes in τ_m over the time course (Fig. 2B; Supplementary Fig. S1D–S1I). Multi-class models of random forest classification were then generated to classify macrophage polarization states [(M0, M(IFNγ), and M(IL4/IL13)] across each of the three experimental timepoints (24, 48, and 72 hours). Classification accuracy of test data predictions for all three groups was highest (90.6%) upon training with 75% and testing on 25% of the total dataset (Fig. 2C). Accuracies of >84% were observed for test data predictions regardless of training/test proportions (Fig. 2C; Supplementary Fig. S1L).

Metabolic inhibitors were used to confirm observed autofluorescence differences between M(IFNγ) and M(IL4/IL13) macrophages. Glycolysis, oxidative phosphorylation, and fatty acid oxidation were inhibited with 2DG, NaCN, and etomoxir, respectively. Fold changes in redox ratio with respect to untreated controls with the same stimulation are shown in Fig. 2D. Inhibition of glycolysis decreased the redox ratio of M(IFNγ) macrophages to a greater degree than M(IL4/IL13) macrophages (Fig. 2D), consistent with published studies that show increased reliance on glycolysis for these antitumor, M1-like macrophages compared with protumor, M2-like macrophages (7). Conversely, M(IL4/IL13) macrophages showed greater increases in redox ratio than M(IFNγ) macrophages with inhibition of oxidative phosphorylation and fatty acid oxidation (Fig. 2D). This is consistent with prior studies that show increased reliance on oxidative phosphorylation and fatty acid oxidation in M2-like macrophages compared with M1-like macrophages (7, 8). Fold changes were statistically significant between control and inhibitor-treated conditions across all treatments (Supplementary Table S2). Fold change of both NAD(P)H and FAD mean lifetimes with inhibitors are shown in Supplementary Fig. S1J and S1K. These findings agree with previous studies reporting upregulation of glycolysis in M1-like macrophages, while M2-like macrophages preferentially shift toward a more oxidative metabolic state (3, 6, 8). Collectively, these results demonstrate that metabolic autofluorescence imaging can assess macrophage phenotype and metabolism.

The variable metabolic behavior observed between polarization states over the time course prompted further analysis of heterogeneity within the aggregated single-cell data. Population density modeling of single-cell redox ratios was performed to determine whether metabolic autofluorescence imaging can resolve macrophage heterogeneity. Population density curves revealed heterogeneous metabolic subpopulations of macrophages across conditions and timepoints (Fig. 3A–C). All stimulation conditions and timepoints had a narrow population of cells with similarly low redox ratio along with a second population of cells with a broad range of higher redox ratios (Fig. 3A–C). Note that the proportion of the high redox ratio subpopulation (p₂) was a substantial proportion of all cells for the M(IFNγ) condition (red) at 48 and 72 hours (Fig. 3B and C), which contributes to its higher average redox ratio at 48 and 72 hours (Supplementary Fig. S1A). In addition, shifts within the low redox population (p₁) and within the high redox population (p₂) between conditions were significant at each timepoint (Fig. 3A–C).These distributions illustrate the heterogeneous metabolic state of macrophages on a single-cell level. Population distributions overlapped between M0, M(IFNγ), and M(IL4/IL13) conditions, which highlights a continuum of metabolic activity between conditions. M(IFNγ) and M(IL4/IL13) macrophages also exhibited greater variability compared with M0 macrophages at 24 and 72 hours (Supplementary Table S3). Increasing numbers of subpopulations were observed from population distribution analysis of single-cell NAD(P)H and FAD τ_m during macrophage stimulation, highlighting substantial heterogeneity in enzyme binding activity (Supplementary Fig. S1M–S1R).

Heterogeneity in macrophage function within stimulation conditions [M(IFNγ) or M(IL4/IL13)] was further assessed from gene and protein expression. Immunofluorescence was quantified from images of M(IFNγ) and M(IL4/IL13) macrophages stained for CD86 and CD206 expression (Fig. 3D). Both stimulation conditions had basal expression of these markers that fluctuated during polarization (36). Thus, the ratio of CD86 to CD206 expression was used to assess protein changes associated with macrophage polarization (37). The CD86 to CD206 ratio increased significantly in M(IFNγ) compared with M(IL4/IL13) conditions by 72 hours, while heterogenous expression was observed at earlier timepoints (Fig. 3D). Gene expression of cytokines and other mediators of inflammation (Supplementary Table S1) was also measured with qRT-PCR as a parallel measure of heterogeneity in macrophage function across 24–72 hours of stimulation (Fig. 3E–F). Ccl2 and Ccl5 expression and downregulation of Il10 and Ptgs2 in M(IFNγ) conditions at 72 hours indicated polarization toward M1-like phenotype (Fig. 3E). IFNγ stimulation at 72 hours also upregulated mixed phenotype Il12b, Il6, Ccl22, and Nos2 and M2-like gene, Stat3. Accordingly, IL4/IL13 stimulation gradually upregulated Il10 and Ptgs2, while Il1b and Vegfa were significantly downregulated by 72 hours, confirming M2-like polarization (Fig. 3F). Variability in expression of M1-like, M2-like, and mixed phenotype genes at 24 and 48 hours of stimulation suggests incomplete polarization at earlier timepoints (Fig. 3E and F).

The Stacks system was previously characterized for diffusion gradients and cell viability, survival, and recruitment (11). To determine nonspecific changes in metabolism, metabolic activity was assessed in 2D breast cancer cultures with and without wavelength mixing (Supplementary Fig. S2A–S2D) and in macrophages 1 hour after seeding for RAW264.7 macrophage monocultures and RAW264.7 + PyVMT mouse breast carcinoma cocultures within the Stacks system (Supplementary Fig. S2E–S2G). Minimal differences in autofluorescence measurements were observed between both imaging methods (Supplementary Fig. S2B–S2D). Autofluorescence measurements were not significantly different between monocultured macrophages and cocultured macrophages, confirming that macrophages were similar for both conditions at the time of seeding (Supplementary Fig. S2E–S2G). In addition, no appreciable depth-dependent attenuation of NAD(P)H and FAD intensities was observed in these studies of optically transparent collagen I gels imaged within approximately 170-μm depths (Supplementary Fig. S2H and S2I), ensuring accurate redox ratio trends without corrections for depth-dependent light attenuation. In addition, RAW264.7 macrophages were seeded on both ends of the ECM layer to account for nonspecific macrophage migration in the 3D Stacks system without tumor stimulation. Minimal change in macrophage migration was detected over 72 hours within macrophage–macrophage cultures (Supplementary Fig. S2J–S2L). Overall, these data suggest that changes in macrophage migration and metabolism in coculture at later timepoints were due to tumor-specific signaling.

Next, metabolic autofluorescence was characterized in microfluidic cultures of mouse PyVMT mammary carcinoma and mouse RAW264.7 macrophages over a 72-hour time course. Redox ratio, NAD(P)H τ_m, and FAD τ_m of cocultured macrophages from a representative z-plane (migration distance = 6 μm) were visually distinct from monocultured macrophages at the same z-plane as early as 24 hours postseeding (Fig. 4A). Morphologic differences also emerged between monocultured and cocultured macrophages over the 72-hour time course. Cocultured macrophages were larger than monocultured macrophages and adopted a mixture of flat, cuboidal, or elongated, spindle-like morphologies, commonly observed for polarized macrophage states (Fig. 4A; ref. 35). Genetic heterogeneity between macrophages within 3D cocultures was evaluated by qRT-PCR for a gene panel associated with M1-like, M2-like, and mixed macrophage phenotypes (Supplementary Table S1; ref. 38). Minimal changes in gene expression of select M1-like, M2-like, and mixed phenotype genes were observed between monocultured and cocultured macrophages at early timepoints (Fig. 4B). However, 72 hours of coculture resulted in significantly increased expression across Ccl2, Ccl5, and Il23a (M1-like), Vegfa and Stat3 (M2-like), and Il6 (mixed phenotype; Fig. 4B). In addition, Il12b was significantly downregulated in cocultured macrophages at 72 hours (Fig. 4B).

Quantitative autofluorescence measurements more clearly captured metabolic changes across 3D macrophage cultures. Significant differences in NAD(P)H τ_m and FAD τ_m were observed between monocultured and cocultured macrophages at 24, 48, and 72 hours, while the redox ratio of cocultured macrophages significantly decreased only at 72 hours compared with monoculture conditions (Fig. 4C–E). The redox ratio gradually decreased over 72 hours of coculture with tumor cells (Fig. 4C), while NAD(P)H and FAD τ_m did not significantly change over time (Fig. 4D and E). Further changes in NAD(P)H and FAD lifetime components were also observed over the coculture time course (Supplementary Fig. S3A–S3F). Heterogeneity during RAW264.7 macrophage polarization in response to PyVMT tumor cell coculture was also quantified with population density modeling of cell-level metabolic autofluorescence over the 72-hour imaging time course (Fig. 4F–H). Two distinct redox ratio subpopulations of macrophages were present in coculture at the 24- and 72-hour timepoints, and in monoculture at 24 and 48 hours (Fig. 4F,–H). Variability in redox ratio of macrophage monocultures was less than cocultures at all timepoints (Supplementary Table S3). Heterogeneity in enzyme binding activity was reflected in single-cell NAD(P)H and FAD τ_m distributions, which also detected macrophage subpopulations during stimulation (Supplementary Fig. S3G–S3L). These results demonstrate that this 3D tumor coculture can stimulate heterogeneous changes in macrophage gene expression (Fig. 4B) that may contribute to the metabolic heterogeneity between individual macrophages that is exclusively captured with metabolic autofluorescence measurements (Fig. 4F–H).

The 3D culture and imaging approach enabled an analysis of migration as a source of metabolic heterogeneity within 3D mouse cocultured macrophages. Specifically, mouse macrophage metabolism was monitored during migration toward the tumor layer within 3D cocultures by acquiring volumes of metabolic autofluorescence images spanning the macrophage layer. Heatmaps of optical redox ratio at each z-plane show that macrophage metabolism and migratory distance vary over the 72-hour time course (Fig. 5A). Heatmaps of NAD(P)H τ_m and FAD τ_m with migration distance are represented in Supplementary Fig. S3M and S3N. The number of macrophages within each z-plane (depth increments of 3 μm) was divided by the total number of macrophages within each image volume to quantify the migration distance of macrophages over time (Fig. 5B–D). Monocultured macrophages exhibited minimal migration from the seeding plane, whereas cocultured macrophages migrated further into the ECM toward the tumor layer especially at 24 and 48 hours (Fig. 5B and C). Therefore, the maximum migration distance for the monoculture condition was used to threshold passive from active macrophage migration at each timepoint (Fig. 5B–D; ref. 32). Actively migrating macrophages in coculture had lower redox ratio at all timepoints, suggesting that these macrophages shift toward oxidative metabolism to enhance migration in the TME (Fig. 5E and F). In addition, NAD(P)H τ_m decreased in actively migrating macrophages at 24 and 48 hours, while FAD τ_m increased significantly at 72 hours (Fig. 5F; Supplementary Fig. S3O and S3P). Z-scores for each metabolic autofluorescence variable were calculated for passively and actively migrating mouse macrophages in coculture by subtracting the variable mean of the monoculture condition from the variable mean per condition, and then dividing by the monoculture SD at each timepoint (Fig. 5F). Multi-class random forest models were generated from all 11 autofluorescence variables to classify passive and active migration in 3D mouse cocultures across all timepoints (24, 48, and 72 hours). High classification accuracy was observed for test data predictions for passive versus active migration, with an accuracy of at least 85% for all test data predictions (Fig. 5G; Supplementary Fig. S3Q). These results highlight the combination of metabolic autofluorescence imaging and 3D Stacks coculture to reveal novel relationships between metabolic activity and macrophage migration.

To further observe human tumor–mediated macrophage polarization and migration, THP-1 human monocytes were cultured alone or with patient-derived IDC cells, and metabolic autofluorescence changes were measured over 72 hours. Qualitative redox images from a representative z-plane (migration distance = 12 μm) highlighted cell-level heterogeneity in redox ratio and lifetimes for cocultures compared with monoculture (Fig. 6A). Increased cell and nucleus size were observed in cocultured cells, consistent with reported morphologic differences between human monocytes and activated macrophages (Fig. 6A; ref. 39). Expression of M1-like and M2-like macrophage signature genes demonstrated heterogeneity in macrophage phenotype for human cocultures (40, 41). Twenty-four hours of coculture significantly upregulated only IL1B, while expression of TNF and PTGS2 significantly decreased in 48-hour cocultures (Fig. 6B). Mixed phenotype genes were variably expressed at both 24 and 48 hours (Fig. 6B). Conversely, mixed phenotype, IL12B, TGFB1, CSF1, NOS2, IL10, was upregulated after 72 hours of coculture with patient-derived tumor cells (Fig. 6B). IL1B and TNF (M1-like) and CCL22, VEGFA, and PTGS2 (M2-like) were also significantly increased by 72 hours (Fig. 6B). Quantitative autofluorescence measurements showed monocyte-derived macrophages in coculture have significantly higher redox ratio at 48 hours (Fig. 6C), no change in NAD(P)H τ_m at any timepoint (Fig. 6D), and significantly lower FAD τ_m at all timepoints compared with monoculture (Fig. 6E). NAD(P)H and FAD lifetime components were also affected by tumor coculture (Supplementary Fig. S4A–S4F). These parallel metabolic and genetic findings demonstrate variation in tumor-mediated monocyte activation and macrophage polarization across species, culture in 2D versus 3D, and breast tumor subtypes, which highlights the unique insights enabled by combining 3D Stacks and autofluorescence imaging technologies.

The considerable heterogeneity observed in metabolic autofluorescence across 2D and 3D models of mouse macrophage polarization prompted an investigation of heterogeneity due to migration within 3D human tumor macrophage cocultures. Here, tumor-mediated migration was also assessed in 3D cocultures of primary IDC and THP-1 monocytes. Heatmaps of redox ratio changes with z-plane show increased migration in coculture conditions and decreased redox ratio in macrophages localized closer to the tumor at 72 hours (Fig. 7A), similar to the behavior of RAW264.7 cells (Fig. 5A). NAD(P)H and FAD τ_m of macrophages also varied with time and z-plane (Supplementary Fig. S4G and S4H). Monocyte-derived macrophages stimulated in coculture exhibited substantial migration toward the tumor layer at all timepoints, in contrast to monocytes in monoculture (Fig. 7B–D). Passively and actively migrating macrophages were then defined on the basis of monoculture migration distances at each timepoint (Fig. 7B–D). Both the redox ratio and FAD τ_m of actively migrating macrophages in coculture were higher at early timepoints compared with passively migrating macrophages, but gradually declined over 72 hours (Fig. 7E and F; Supplementary Fig. S4J). Actively migrating macrophages in coculture also had decreased NAD(P)H τ_m at 24 and 72 hours (Supplementary Fig. S4I). Z-scores for metabolic autofluorescence of passively and actively migrating human macrophage cocultures were calculated relative to the monoculture condition at each timepoint, demonstrating substantial differences between monocultured and actively migrating cocultured macrophages, while passively migrating cocultured macrophages minimally changed relative to monoculture (Fig. 7F). Random forest classification with all 11 autofluorescence variables was also performed to distinguish passively and actively migrating macrophages at 24, 48, and 72 hours of coculture with human primary invasive ductal carcinoma. Classification accuracy for passive and active migration was >83% across all test data predictions (Fig. 7G; Supplementary Fig. S4K). Lastly, the versatility of this platform was demonstrated from a time course of autofluorescence changes in THP-1 human monocytes in coculture with a triple-negative breast cancer cell line (MDA-MB-231; Supplementary Fig. S5A–S5J) in contrast with patient-derived IDC cocultures (Figs. 6 and 7), highlighting breast cancer cell origin as a source of variability in monocyte-derived macrophage metabolism. Overall, these results establish the utility and versatility of the platform to address the challenge of evaluating dynamic macrophage function across human TME models.

Macrophage plasticity leads to frequent shifts in function and metabolism in the TME, resulting in heterogeneous tumor-associated macrophage populations (9). Here, we show that metabolic autofluorescence imaging can monitor macrophage heterogeneity during polarization and migration within a 3D microfluidic model of the TME. Metabolic autofluorescence imaging enables nondestructive monitoring of metabolic changes in individual, live cells within 3D in vitro cultures and in vivo tumors (16–18, 22). In addition, 3D microfluidic culture systems, such as the Stacks microfluidic device used in this study, provide simple platforms to recapitulate dynamic environmental conditions that characterize in vivo tumors (e.g., hypoxia, acidosis, and nutrient starvation) and are powerful to study primary human cells in an in vivo–like environment (42). Previous studies have shown that metabolic autofluorescence can identify macrophages in 2D cultures and in vivo, but have not explored spatiotemporal heterogeneity in macrophage metabolism within the TME (16–18, 24). This is the first study to monitor cell-level macrophage metabolism in microfluidic models of intact TME. Here, we quantify metabolic heterogeneity of macrophages in the Stacks 3D microdevice platform using autofluorescence imaging to provide novel insights into spatiotemporal heterogeneity of macrophages in the TME.

Temporal changes in microenvironment conditions affect macrophage metabolism and function, so this study quantified time-dependent changes in macrophage autofluorescence within microfluidic 3D tumor cocultures. This study first confirmed that metabolic autofluorescence imaging can quantify changes in macrophage metabolism and distinguish macrophage subpopulations within intact, living samples. Differences in NAD(P)H and FAD autofluorescence were first observed in 2D cytokine-stimulated macrophages over 72 hours and reflected the expected metabolic shifts for M(IFNγ) and M(IL4/IL13) macrophages (Fig. 2; refs. 43–45). This result is consistent with previous metabolic flux and metabolite accumulation studies showing increased glycolysis in LPS- or IFNγ + TNFα–stimulated (M1-like) macrophages and increased tricarboxylic acid cycle in IL4-stimulated (M2-like) macrophages, as well as NAD(P)H lifetime studies distinguishing IFNγ/LPS-treated and IL4/IL13-treated mouse bone marrow–derived macrophages (16). Generation of a multiparametric classification model incorporating all 11 metabolic autofluorescence measurements further improved discrimination of these 2D macrophage polarization states (Fig. 2C; Supplementary Fig. S1L). The classifier performed well for identifying M0, M(IFNγ), and M(IL4/IL13) macrophage conditions regardless of stimulation period, indicating that this classifier could accurately identify polarization in samples where the period of stimulation is unknown. In addition, metabolic heterogeneity was observed across population density distributions of single-cell redox ratios for 2D mouse macrophages over 24–72 hours of polarization, revealing distinct populations of low and high redox ratio (Fig. 3A–C). This metabolic heterogeneity is consistent with heterogeneous gene and protein expression profiles (Fig. 3D–F), but autofluorescence measurements resolve cell-level heterogeneity, unlike these bulk gene and protein measurements. This analysis also demonstrates the sensitivity of autofluorescence measurements to temporal functional plasticity commonly reported during macrophage polarization. For example, the bimodal populations observed in redox ratio of polarized macrophages may reflect early- and late-phase metabolic reprogramming as shown in recent studies to promote initial upregulation of glycolytic or oxidative processes prior to reestablishing a resting metabolic state (43, 46). Notably, autofluorescence imaging provides a unique tool to monitor metabolically distinct cell subpopulations, unlike standard metabolic flux or metabolite production measurements that monitor bulk changes within a pooled population of cells.

Next, RAW264.7 macrophages cocultured in 3D with PyVMT breast cancer displayed changes in redox ratio consistent with metabolic changes from tumor stimulation in previous studies (Fig. 4A and C; ref. 47). Differences in redox ratio and fluorescence lifetime measurements were observed between RAW264.7 macrophages in cytokine-stimulated 2D cultures (Fig. 2B) and tumor-stimulated 3D cultures (Fig. 4C–E), consistent with reported differences in intracellular metabolite concentration and metabolic flux based on stimulation condition (e.g., cytokines and tumor conditioned media) and culture in 2D versus 3D (48, 49). In addition, 3D cocultured RAW264 macrophages exhibited heterogeneous cell-level autofluorescence and pooled gene expression (Fig. 4B and F–H). This is consistent with reported time courses of metabolic and genetic changes in cocultures of tumor cells and monocytes or macrophages (50). Concentrations of metabolites, including NAD(P)H and FAD, can be replenished or depleted more rapidly than the downstream gene expression initiated by these metabolic changes, so correlations between gene expression and autofluorescence changes were difficult to interpret (50). However, the genetic and metabolic heterogeneity observed here is supported by recent studies that similarly show tumor-associated macrophages adopt a mixture of M1-like and M2-like characteristics and functions (51). More importantly, autofluorescence measurements capture early metabolic changes that are followed by these late changes in gene and protein expression (52). These different timescales of metabolic, gene, and protein expression changes in macrophages are likely responsible for differences between autofluorescence measurements and standard genomic/proteomic readouts. This reinforces a model of tumor-associated macrophages with diverse phenotypes (53), which can be further explored using the methods demonstrated here. Ultimately, these studies show that metabolic autofluorescence imaging is sensitive to heterogeneity in macrophage function and metabolism within 3D models of the TME.

Microfluidic culture is also attractive for mimicking the 3D spatial structure of the TME. Accordingly, multi-photon excitation of metabolic autofluorescence enables cellular resolution imaging in thick, scattering samples to monitor environmental gradients and spatial heterogeneity in 3D tumor models (54). Therefore, we measured metabolic heterogeneity of RAW264 macrophages with 3D migration during coculture with PyVMT tumor cells (Fig. 5). Macrophages in coculture exhibited greater migration across the ECM layer over 72 hours compared with monocultured macrophages, which was shown to be a tumor-specific behavior (Fig. 5A–D; Supplementary Figs. S2J–S2L and S3M and S3N). Previous studies suggest that mouse macrophages often exhibit random migration during chemotaxis (55). This is consistent with our observations of mouse macrophage migration (Fig. 5A–D). Prior studies have also shown that M2-like macrophages are typically more migratory, which is consistent with the oxidative metabolic shift in mouse macrophages localized closest to the tumor layer (actively migrating; Fig. 5E; refs. 56, 57). However, variability in individual metabolic autofluorescence measures [e.g., redox ratio and NAD(P)H/FAD lifetimes] limits their classification accuracy when used alone. Thus, classification models based on all 11 metabolic autofluorescence variables were generated to collectively describe metabolic changes in macrophages with tumor stimulation in 3D and improve discrimination of passive and active migratory phenotypes (Figs. 5G and 7G; Supplementary Figs. S3Q and S4K). Notably, this 3D imaging platform enables further exploration of cell-level metabolic heterogeneity as a function of migration, not resolved from the genetic profile pooled across all migratory depths. Overall, metabolic autofluorescence imaging and microfluidic models of 3D microenvironment illustrate tumor-driven spatiotemporal changes in macrophage metabolism.

An important advantage of the Stacks system is the use of primary human cells for studies of human tumor–immune interactions. We highlighted this advantage with 3D cocultures of primary IDC and THP-1 monocytes (Fig. 6). Previous studies have shown that human THP-1 monocytes will differentiate and preferentially polarize to M2-like macrophages as early as 48 hours after treatment with tumor-conditioned media (58, 59). This is consistent with redox ratio changes of THP-1s in coculture with primary tumor cells (Fig. 6C). The late oxidative, M2-like switch captured by the redox ratio is also consistent with previous reports of 2D-differentiated THP-1 cells exposed to IL4 (M2-like) cytokine stimulation (18). Although autofluorescence changes were less pronounced in human cocultures when all macrophages were pooled, spatial analyses revealed more substantial autofluorescence changes due to macrophage migration in coculture. Cocultured THP-1s also migrated farther across the ECM than monocultures over 72 hours (Fig. 7A–D). Human macrophages and monocytes exhibit strong directional migration toward chemotactic gradients (Fig. 7B–D), consistent with our observations of human macrophage and monocyte migration (12). The redox ratio of actively migrating human macrophages increased at 24 hours and gradually declined over 72 hours (Fig. 7E and F), consistent with reports of upregulated glycolytic metabolism prior to oxidative shifts in human macrophages during chemoattractant-induced migration (60). These studies indicate that metabolic autofluorescence imaging in human tumor macrophage microfluidic cultures provides a valuable platform to characterize novel spatiotemporal dynamics of macrophage metabolism within the human TME.

Cell function is variable across species and individuals. This variability is reflected in metabolic autofluorescence differences between mouse PyVMT/RAW264.7 and human primary invasive ductal carcinoma/THP-1 cocultures (Figs. 4 and 6; Supplementary Figs. S3 and S4). Our results are consistent with previous reports of genetic and metabolic differences between mouse and human macrophage lines, including RAW264.7 and THP-1s (40, 41, 43, 44, 61). These changes are likely due to distinct profiles of cytokine secretion and other polarizing stimuli (36, 40). Differences in macrophage response to tumor coculture are also likely due to unique tumor cell origin and functional activity. The human primary invasive ductal carcinoma line was derived from resected tissue of a moderately differentiated and hormone receptor–positive breast cancer, while PyVMT mouse breast carcinoma originates from a transgenic mouse model with spontaneous oncoprotein-controlled tumor growth that more closely mimics human triple-negative breast cancer (25, 62). Furthermore, genetic and transcriptional profiles of human breast cancer are only partially conserved in PyVMT mouse mammary cancer (62). Therefore, variations in macrophage metabolism in response to these breast cancer cocultures are likely due to differences in cell phenotype and function (e.g., mutation/receptor status and secretory profiles) across species.

The combined metabolic autofluorescence imaging and 3D Stacks microscale platform provides a unique system to evaluate tumor–macrophage cross-talk within the TME. Although this model system mimics key components of the TME, limitations exist for fully recapitulating the cellular and architectural diversity of the TME (10). One limitation in our studies is the use of macrophage cell lines rather than primary macrophages. These macrophage cell lines (RAW264.7 and THP-1) were previously characterized under conditions similar to those tested in this study (11, 43, 47), so these cell lines were chosen to provide a validated model for autofluorescence imaging in 3D Stacks. However, RAW264.7 and THP-1 lines originated from monocytic leukemias, which introduces phenotypic and functional differences related to the origin site and malignant background compared with primary macrophages derived from peripheral blood or bone marrow (63, 64). The autofluorescence imaging and 3D Stacks system demonstrated here are compatible with primary macrophages, which should be investigated in the future to provide deeper insights into functional dynamics in macrophage migration and metabolism. Furthermore, environmental factors (e.g., oxygen and pH gradients) are another source of contrast for autofluorescence lifetimes (19, 42). Decreased FAD autofluorescence lifetimes have been previously correlated with increased pH, which suggests that our observed decreased FAD mean lifetimes with prolonged coculture and increased migration in human monocyte–derived macrophages may be due increased local pH levels (65). An additional consideration for this platform is the efficiency of excitation using the wavelength mixing approach. This method has been shown to excite FAD more efficiently than two-color sequential imaging (27). However, wavelength mixing and two-color sequential imaging provide comparable excitation efficiency for NAD(P)H and comparable fluorescence lifetime values for both fluorophores (27). In our study, all redox ratio values were referenced to control groups, and relative changes in redox ratio were conserved between wavelength mixing and two-color sequential imaging (Supplementary Fig. S2). Consideration of these constraints for future adaptations of this 3D imaging and microscale culture platform should improve its versatility for characterizing cellular dynamics in the TME.

Overall, we have established a novel, single-cell imaging and 3D microfluidic culture platform to monitor the spatial and temporal dynamics of macrophage metabolism within the TME. Autofluorescence imaging of 3D tumor–macrophage cocultures characterized metabolic heterogeneity with tumor-mediated macrophage polarization and migration. This approach could be used to evaluate metabolic heterogeneity in more complex 3D microscale cultures with additional immune cells, stromal cells, and vasculature. In addition, this technology supports live, high-throughput biological studies to probe mechanisms of macrophage response to microenvironmental stimulation in parallel to imaging studies. Direct manipulation of environmental gradients (e.g., oxygen, nutrients, and pH) during culture and imaging could assess the relationship between environmental pressures and macrophage metabolism, function, and organization. Ultimately, these tools may improve our understanding of cellular heterogeneity and metabolism in the TME, and their effects on tumor progression and treatment response.

T.M. Heaster reports grants from National Science Foundation during the conduct of the study. J. Yu reports a patent for US20170239661 issued. M.C. Skala reports grants from NIH and National Science Foundation during the conduct of the study. No disclosures were reported by the other authors.

T.M. Heaster: Conceptualization, resources, data curation, software, formal analysis, validation, investigation, visualization, methodology, writing-original draft, project administration, writing-review and editing. M. Humayun: Resources, data curation, formal analysis, validation, investigation, methodology, writing-original draft, writing-review and editing. J. Yu: Conceptualization, resources, validation, investigation, methodology, writing-review and editing. D.J. Beebe: Conceptualization, resources, supervision, writing-review and editing. M.C. Skala: Conceptualization, resources, supervision, funding acquisition, writing-original draft, writing-review and editing.

We would like to thank Rupsa Datta and Jose Ayuso for their valuable input on the experimental design and article composition. We acknowledge Jens Eickoff for his guidance in statistical analysis of the reported data. We also acknowledge Kayvan Samimi and Matthew Stefely for input on design of scientific diagrams and data representation. M.C. Skala acknowledges support from the NCI (R01 CA211082, R01 CA205101, and R01 CA185747), the NSF (CBET-1642287), Stand Up to Cancer (SU2C-AACR-IRG-08-16 and SU2C-AACR-PS-18), the Morgridge Institute for Research, and the University of Wisconsin Carbone Cancer Center. Stand Up To Cancer is a division of the Entertainment Industry Foundation. The cited SU2C grants were administered by the American Association for Cancer Research, the scientific partner of SU2C. T.M. Heaster acknowledges the support of the NSF Graduate Research Fellowship (DGE-1256259). M. Humayun acknowledges support by NIH fellowship F31 CA247248.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.